Explain an ML project end-to-end with tradeoffs

Choose one ML project you have shipped to production and describe it from start to finish. Be precise: 1) How you framed the problem (prediction versus causal decision-making), what the target was, and how you avoided label leakage; 2) Which data sources you used, the sampling window, and the offline metric(s) along with your reasoning (for instance, AUC versus calibration/Brier when the goal is monetization); 3) Your feature engineering, treatment of sparse or categorical signals, and how you enforced privacy and fairness constraints; 4) The models you considered and their trade-offs (such as XGBoost versus shallow neural nets versus GLM), your hyperparameter approach, and any ablation studies you ran; 5) How you analyzed errors and monitored after deployment (drift, stability, guardrail metrics); 6) How you turned model improvements into product impact when an A/B test wasn't available (for example, causal uplift modeling, CUPED, or backtests); 7) What you would change in a second version if you had twice the data or tighter latency constraints. Overview: This question assesses a candidate's ability to design and deliver an end-to-end machine learning system, touching on problem framing, target definition and label leakage prevention, data and metric selection, feature engineering under privacy and fairness constraints, model trade-offs, hyperparameter and ablation analysis, and post-deployment monitoring and impact measurement. It is often used to gauge hands-on production experience and trade-off reasoning in machine learning, testing both applied skills and conceptual knowledge of modeling, evaluation, and operational limits.